Guojin Si

The economic viability of offshore wind farms (OWFs) is heavily constrained by exorbitant operation and maintenance (O&M) costs and significant revenue losses during system downtime. While opportunistic maintenance (OM) strategies have been widely adopted to share the high fixed costs of vessel dispatch, existing frameworks largely rely on static degradation models and fail to account for the extreme short-term volatility of modern deregulated electricity markets. To bridge these gaps, this paper proposes a dynamic OM framework driven by day-ahead locational marginal prices (LMPs). Firstly, a hybrid multi-component degradation model is established. A stochastic process is developed to capture the complex crack initiation and propagation mechanisms of wind turbine blades, while the progressive aging of electromechanical components is characterized by Weibull distributions. Secondly, a dynamic and price-responsive OM thresholding mechanism is introduced. By integrating day-ahead weather forecasts and LMPs, the proposed strategy proactively groups preventive maintenance tasks during low-price windows and tightens thresholds during price spikes, thereby minimizing the expected maintenance cost. Numerical case studies based on real-world OWF configurations and historical market data demonstrate that the proposed dynamic strategy significantly outperforms traditional static OM approaches, reducing the expected lifecycle maintenance costs by 18.4% and substantially mitigating financial risks associated with electricity price volatility.

With the increasing complexity and scale of modern equipment systems, remaining useful life (RUL) prediction faces three major challenges: (1) feature redundancy in multi-source sensor data, (2) difficulty in modeling long-term temporal dependencies in degradation processes, and (3) multi-modal degradation patterns arising from multiple failure modes. To overcome these challenges, this study proposes a novel prediction framework that integrates dual-attention temporal convolutional networks with a selective ensemble strategy driven by particle swarm optimization (DATCN-PSOSEN). Specifically, the model incorporates channel attention at the input end of the TCN structure, enabling the model to adaptively evaluate the importance of sensor channels. Simultaneously, the temporal attention mechanism applies weighting to the time dimension of features extracted by the TCN, enhancing the model’s ability to dynamically capture temporal dependencies, thus forming the DATCN base learner. Multiple sliding window lengths are used to diversify base learners. Additionally, the PSOSEN employs the PSO algorithm to effectively enhance the model’s robustness by automatically pruning and weighting base learners. Extensive experiments on the C-MAPSS dataset show that the proposed DATCN-PSOSEN achieves superior overall performance compared to state-of-the-art RUL prediction models. Notably, it reduces RMSE by up to 14.2% and Score by 24.7% relative to a typical CNN-based baseline (DCNN).

Effective fault diagnosis for industrial robots is critical to ensure their operational availability and reliability in safety-critical applications. Proprioceptive signals from robot servo drive systems provide a nonintrusive and promising solution for practical in situ diagnosis. While their exhibited significant multisensor heterogeneity and nonstationarity caused by time-varying operating conditions challenge existing diagnostic methods in identifying fault-induced intrinsic signatures. Concurrently, inadequate prior knowledge integration limits data-driven approaches in physical consistency, leading to outputs that may not adhere to fundamental physical principles, such as the laws of motion derived from Lagrangian mechanics. Thereupon, a hybrid attention Lagrangian variational autoencoder with latent representation disentanglement is proposed for industrial robot diagnosis under time-varying conditions using multivariate proprioceptive signals. Specifically, a hybrid attention-based encoder is devised for composite modeling of interseries association patterns and intraseries temporal variations, encoding multimodal spatiotemporal complementary information into condition-independent latent spaces. A Lagrangian mechanics-embedded decoder with structured network topology ensures latent spaces with critical information reservation and nonstationary disturbance isolation by physics-constrained state reconstruction. Concurrently, the dual-regularized latent disentanglement strategy separates fault-specific latent space from fault-invariant contexts through the variational evidence lower bound-driven posterior alignment and the explicit supervision from adversarial and diagnostic classifiers. Hereinto, the Gaussian mixture prior is imposed on the fault-specific latent space to promote interclass separability and intraclass compactness. The developed physics-embedded framework effectively decouples intrinsic fault patterns from nonstationary inference. Comprehensive case studies using both simulated and real-world industrial robot datasets are conducted to demonstrate the superiority of the proposed method.